Adaptive Fine-Grained Attention and Multi-Scale Fusion Mechanism for Real-Time Small Traffic Object Detection

Abstract

Unmanned Aerial Vehicle (UAV) video surveillance has become indispensable for intelligent transportation systems, autonomous driving and unmanned vehicle applications. However, traffic objects captured from high altitudes typically exhibit small sizes and weak feature representation, while being susceptible to interference from UAV motion and complex road environments, resulting in detection challenges. Additionally, multi-scale object detection requires a balance between accuracy and efficiency. To address these issues, we propose a real-time small traffic object detection method based on adaptive fine-grained attention and multi-scale fusion mechanism. Building upon the Real-Time Detection Transformer (RT-DETR), we construct a Small Object Enhanced Real-Time Detection Transformer (SOE-RTDETR) with an adaptive fine-grained channel attention block to enhance small object feature extraction. The SOE-RTDETR incorporates CSP-Dilated Reparam Residual Blocks (CSP-DRRB) that expand the convolutional receptive field while maintaining computational efficiency through depthwise convolution and reparameterization. We further optimize the attention-based intra-scale feature interaction module using deformable attention mechanism and propose a bidirectional feature pyramid network specifically designed for small traffic object detection to strengthen multi-scale feature fusion. Experimental results on the VisDrone2019 dataset demonstrate that the proposed SOE-RTDETR achieves improvements of 2.9% in mAP@50 and 2.3% in mAP@50:95 on the validation set, and 2.3% and 1.5% on the test set compared to the baseline RT-DETR-r18 model, while maintaining equivalent model size and computational cost.